AWS described fine-tuning GPT-OSS-20B for multilingual reasoning
AWS published a procedure for fine-tuning GPT-OSS-20B using SFT and GRPO on Amazon SageMaker HyperPod. In the test conducted by the authors, the base model followed the requested reasoning language in only 22 % of 50 examples.
AWS described a procedure for fine-tuning GPT-OSS-20B intended to improve adherence to the requested language when reasoning and answering. The authors tested ten mathematical and logical tasks in five languages: English, French, German, Spanish and Italian. According to their measurements, the base model used the correct reasoning language in 11 out of 50 cases, or 22 %. It achieved 100 % for English, 10 % for German and 0 % for the other languages.
The procedure uses the HuggingFaceH4/Multilingual-Thinking dataset with approximately 949 training and 50 validation examples. The authors used the PyTorch TRL library and the LoRA method for initial training with SFT and GRPO. According to the authors, SFT teaches the desired behavior, while GRPO favors better responses from a group of outputs and can refine behavior the model already sometimes manages to perform.
Training took place on Amazon SageMaker HyperPod with orchestration through Amazon EKS. The described SFT run had 1 000 steps and used four ml.g6e.12xlarge instances, totaling 16 L40S GPUs with 48 GB of memory per GPU. The configuration included FSDP, bf16 precision and gradient checkpointing; Amazon FSx for Lustre stored the data and checkpoints. Details are available in the source article.
Why it matters
The guide provides a concrete starting procedure for teams that need to control the reasoning language of a model serving non-English-speaking users. The baseline measurement also shows why language adherence should be checked separately. However, the result from 50 examples alone does not demonstrate general performance or the benefit of subsequent fine-tuning.
Two audiences, two different impacts
What this means
For individuals
Developers get a procedure for testing adherence to the requested language and preparing to fine-tune GPT-OSS-20B using SFT and GRPO.
For a business
For a team planning its own fine-tuning, the scale of the computing infrastructure matters: the example SFT run used 16 GPUs and shared storage. This configuration provides a basis for capacity planning, rather than a demonstrated minimum requirement.
DevelopmentCheck the original
Event sources
only one source so far · 1 publisher, 0 independent. We count feeds from the same owner only once.